Hidden in Plain Sight: Advancing Harmful Content Understanding via Perception-Aware Adversarial Self-Play

Key points

• Despite significant advances in language understanding, Large Language Models (LLMs) can still miss meaning hidden in plain sight, failing to identify harmful intent in expressions readily understood by humans.1
• Malicious users exploit this gap by strategically disguising harmful content to evade automated moderation, posing a serious threat to online platform safety.1
• The vulnerability reflects a fundamental cognitive asymmetry between human readers, who integrate contextual, phonetic, and visual cues to decipher the intent behind deliberate obfuscation, and text-based LLMs, which rely primarily on token sequences and have limited exposure to evolving evasions.1
• Extensive experiments show that PASP reduces adaptive attack success rates by up to 73.71% and 74.49% on ChineseHarm-Bench and EvoHarmBench, respectively, preserves model capabilities with minimal degradation, and effectively interprets real-world evasive expressions in e-commerce.1

Sources (1)

• [1]Hidden in Plain Sight: Advancing Harmful Content Understanding via Perception-Aware Adversarial Self-Play
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 11, 08:50 AM

  To bridge this asymmetry, we propose PASP (Perception-Aware Adversarial Self-Play), which combines multi-perceptual views with adversarial self-play to improve understanding of harmful intent as evasions evolve.

  Despite significant advances in language understanding, Large Language Models (LLMs) can still miss meaning hidden in plain sight, failing to identify harmful intent in expressions readily understood by humans.

Extractive summary: sentences quoted from the sources.